A federated learning system and method for autonomous driving vehicle networking

Abstract

The application discloses an automatic driving vehicle networking federated learning system which comprises a local training unit, a privacy protection unit, an intermediate layer aggregation unit and a global aggregation layer unit; wherein the local training unit is connected with the intermediate layer aggregation unit through a wireless channel and executes the privacy protection unit; the intermediate layer aggregation unit executes the privacy protection unit and is connected with the global aggregation layer unit through a wireless channel. The automatic driving vehicle networking federated learning system reduces the risk of the updated model parameters being stolen by malicious attackers, improves the security of the federated learning, protects the vehicle privacy information from being stolen and enhances the performance of the automatic driving vehicle.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicle networks, and more particularly to a federated learning system and method for autonomous vehicle networks. Background Technology

[0002] With economic and social development, the cost of automobiles has significantly decreased, and people's incomes have been continuously increasing, leading to an exponential increase in vehicle ownership. Clearly, vehicles have greatly facilitated our lives. However, the increase in vehicles has also brought about an increase in traffic accidents and a higher rate of traffic system paralysis. Therefore, improving the safety and efficiency of the traffic system has become a top priority. Autonomous Vehicular Networks (AVNs) enable collaborative decision-making and control among multiple autonomous vehicles (AVS) to achieve a high level of autonomous driving and a greater degree of intelligent transportation systems. In addition, reducing carbon dioxide emissions, reducing traffic accidents, and improving traffic flow are also some of the advantages brought by AVNs. In AVNs, the realization of autonomous driving requires the coordination of multiple modules such as perception, decision-making, and control. The operation of these modules relies on models trained by artificial algorithms. During the training process, a large amount of data is used, and this data is highly sensitive, including location information, traffic conditions, and the privacy information of vehicle owners. Generally speaking, this information is easily stolen and modified. Furthermore, each vehicle generates a large amount of data every day. Uploading all this data to a cloud server (Elastic Computer Service, ECS) would result in expensive communication costs. Federated learning has solved these problems to some extent.

[0003] Federated learning is a distributed machine learning approach that allows multiple stakeholders to collaboratively train a global model while protecting data privacy. It's commonly used in AVNs (Automatic Vehicle Networks) to implement autonomous driving. Specifically, the AI ​​model is trained on local data within the AV, and only information containing model updates is sent to a central server for aggregation. This approach effectively addresses data privacy and security issues and significantly reduces resource consumption (e.g., energy, bandwidth). However, while federated learning can benefit autonomous driving in AVNs, it still faces several security and privacy threats. First, the updated model parameters could be used by the server to launch inference attacks against the AV. Second, the model parameters transmitted by the AV could be eavesdropped on by attackers due to poor communication, leading to privacy breaches. Finally, inappropriate round values ​​can affect algorithm performance and hinder model convergence. Summary of the Invention

[0004] In view of the aforementioned shortcomings of the prior art, the technical problem to be solved by the present invention is that model parameters in existing autonomous driving vehicle-to-everything (V2X) federated learning systems are easily stolen by malicious attackers, resulting in low security performance and poor model performance. The present invention provides a federated learning system for autonomous driving V2X, reducing the risk of updated model parameters being stolen by malicious attackers, improving the security of federated learning, protecting vehicle privacy information from theft, and enhancing the performance of autonomous vehicles.

[0005] This invention provides a federated learning system for autonomous vehicle networking, comprising a local training unit, a privacy protection unit, an intermediate aggregation unit, and a global aggregation layer unit; wherein, the local training unit executes the privacy protection unit, the local training unit is connected to the intermediate aggregation unit via a wireless channel, the intermediate aggregation unit executes the privacy protection unit, and the intermediate aggregation unit is connected to the global aggregation layer unit via a wireless channel.

[0006] Furthermore, the local training unit is responsible for performing local training in federated learning, including the autonomous vehicles participating in federated learning; the privacy protection unit is responsible for protecting the privacy of the local model parameters of the local training unit and the intermediate model parameters of the intermediate layer aggregation unit; the intermediate layer aggregation unit is responsible for performing intermediate layer model aggregation in federated learning, including the Road Side Unit (RSU); the global aggregation layer 400 is responsible for performing global model aggregation in federated learning and reducing the number of intermediate layer aggregations and global aggregations, including the base station.

[0007] Furthermore, the privacy protection unit is implemented using a differential privacy algorithm, and the local training unit and the intermediate layer aggregation unit also execute the differential privacy algorithm to ensure that the federated learning system has higher security.

[0008] Furthermore, the global aggregation layer unit reduces the number of intermediate layer aggregations and global aggregations based on training efficiency, and broadcasts this information to the intermediate layer aggregation unit and the local training unit to ensure training efficiency.

[0009] Another preferred embodiment of the present invention provides a federated learning method for autonomous vehicle networking, comprising the following steps:

[0010] S100. Autonomous vehicles containing local datasets perform local training in federated learning to obtain local model parameters;

[0011] S200: The autonomous vehicle executes a privacy protection algorithm on the local model parameters and sends the noisy local model parameters to the intermediate layer aggregation unit via a wireless channel;

[0012] S300: Each intermediate layer aggregation unit performs intermediate layer training in federated learning on the received local model parameters to obtain an intermediate model;

[0013] S400: Each intermediate layer aggregation unit performs a privacy protection algorithm on the intermediate model parameters and sends the parameters to the global aggregation layer unit via a wireless channel;

[0014] The S500 global aggregation layer performs global aggregation in federated learning on the received intermediate model parameters to obtain global parameters. Based on the training efficiency, it adjusts the training epochs and outputs all model parameters and the number of training epochs at this time.

[0015] Furthermore, after step S500, step S500A is also included: determining whether the model performance meets the requirements or whether the number of iterations exceeds the set value. If yes, the process ends; otherwise, the process returns to step S100 to continue iterating the global model.

[0016] Furthermore, S100, the autonomous vehicle containing the local dataset, performs local training in federated learning to obtain local model parameters, specifically including the following steps:

[0017] Step 110: The local training unit of the autonomous vehicle samples the sample group to obtain the dataset sample set;

[0018] Step 120: The local training unit updates each sample in the obtained sample set locally;

[0019] Step 130: The local training unit prunes the gradient of each updated sample;

[0020] Step 140: The local training unit updates the local model based on the local gradient obtained after pruning; then, it is determined whether the number of iterations has reached the local iteration count condition. If yes, proceed to step 150; otherwise, return to step 110 to continue local model training.

[0021] Step 150: The local training unit performs a privacy protection unit on the local model parameters to obtain local model parameters with added noise.

[0022] Furthermore, S400 and each intermediate layer aggregation unit perform privacy protection algorithms on the intermediate model parameters and transmit the parameters to the global aggregation layer unit via a wireless channel, specifically including the following steps:

[0023] Step 410: Obtain the intermediate layer model parameters; determine if the number of iterations has reached the intermediate iteration count condition. If yes, proceed to step 420A; otherwise, proceed to step 420B.

[0024] Step 420A: Each intermediate aggregation layer unit broadcasts the aggregated parameters to the autonomous vehicles in the area;

[0025] Step 420B: Each intermediate layer performs a privacy protection algorithm on the intermediate layer model parameters to obtain intermediate layer model parameters with added noise;

[0026] Step 430: Compare the benefits of training each intermediate layer with the intermediate model benefit threshold to obtain the number of RSUs whose training benefits are less than the threshold;

[0027] Step 440: Send the noisy intermediate layer model parameters and the number of RSUs with training benefits less than the threshold to the global aggregation layer via a wireless channel.

[0028] Furthermore, the S500 and global aggregation layer perform global aggregation in federated learning on the received intermediate model parameters to obtain global parameters. Based on the training efficiency, the training epochs are adjusted, and all model parameters and the number of training epochs are output. Specifically, this includes the following steps:

[0029] Step 510: The global aggregation layer unit aggregates the intermediate layer model parameters sent in to obtain the global model parameters;

[0030] Step 520: The global aggregation layer unit reduces the number of aggregations in the intermediate aggregation layer based on the RSU ratio of the intermediate layer training benefit that is less than the threshold sent by the intermediate layer aggregation unit.

[0031] Step 530: The global aggregation layer unit determines the relationship between the benefits of global training and the global model benefit threshold, thereby reducing the number of global aggregations.

[0032] Step 540: The global aggregation layer unit broadcasts the aggregated parameters to each autonomous vehicle.

[0033] Step 550: The global aggregation layer unit broadcasts the new aggregation count of the intermediate layer aggregation unit and the global aggregation layer to all intermediate layer aggregation units and all autonomous vehicles.

[0034] Technical effect

[0035] This invention discloses a federated learning system and method for autonomous vehicle networking. To address the privacy leakage issue that still exists during data transmission in federated learning, a privacy protection unit is established. Before transmitting data via wireless channel, both the autonomous vehicle and the global aggregation layer unit perform a differential privacy algorithm on the model parameters. This defends against inference attacks launched by the server against the updated model parameters and against eavesdropping attacks by attackers, effectively protecting vehicle privacy.

[0036] In implementing differential privacy within a federated learning environment, there exists an optimal number of communication rounds. To approach this optimal number of rounds, this invention achieves better results by adjusting the number of communication rounds during training using a discount method. By reducing the number of training rounds, superior model performance can be obtained within a limited number of rounds.

[0037] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of a federated learning system for autonomous vehicle networking, according to a preferred embodiment of the present invention.

[0039] Figure 2 This is a flowchart illustrating a preferred embodiment of the federated learning method for autonomous vehicle networking according to the present invention.

[0040] Figure 3 This is a schematic diagram of the training process of a local training unit in a federated learning method for autonomous vehicle networking according to a preferred embodiment of the present invention.

[0041] Figure 4 This is a schematic diagram of the operation flow of the privacy protection unit of a federated learning method for autonomous vehicle networking according to a preferred embodiment of the present invention.

[0042] Figure 5 This is a schematic diagram of the training process of the intermediate layer training unit of a federated learning method for autonomous vehicle networking according to a preferred embodiment of the present invention.

[0043] Figure 6 This is a schematic diagram of the training process of the global aggregation layer unit of a federated learning method for autonomous vehicle networking according to a preferred embodiment of the present invention. Detailed Implementation

[0044] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0045] In the following description, specific details, such as particular internal procedures and techniques, are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of the invention with unnecessary detail.

[0046] like Figure 1As shown, a preferred embodiment of the present invention provides a federated learning system for autonomous vehicle networking, including a local training unit 100, a privacy protection unit 200, an intermediate layer aggregation unit 300, and a global aggregation layer unit 400; wherein, the local training unit 100 executes the privacy protection unit; the local training unit 100 is connected to the intermediate layer aggregation unit via a wireless channel; the intermediate layer aggregation unit 300 executes the privacy protection unit 200; and the intermediate layer aggregation unit 300 is connected to the global aggregation layer unit 400 via a wireless channel.

[0047] The local training unit 100 is responsible for performing local training on vehicles that possess local training sets, including autonomous vehicles participating in federated learning. Autonomous vehicles train their local models based on the local dataset and the global model parameters from the previous round to obtain local model parameters. In terms of communication, autonomous vehicles communicate with the RSU via a wireless channel, and also communicate with each other using dedicated short-range communication.

[0048] The privacy protection unit 200 is responsible for encrypting and protecting the model parameters of the local training unit 100 and the intermediate layer aggregation unit 300. Specifically, the privacy protection unit includes an algorithm, namely a differential privacy algorithm. When the local iteration count reaches the local iteration count condition, the local training unit 100 performs the differential privacy algorithm on the local model parameters to obtain noisy local model parameters. Similarly, when the intermediate layer iteration count reaches the intermediate layer iteration count condition, the intermediate layer aggregation unit performs the differential privacy algorithm on the intermediate layer model parameters to obtain noisy intermediate layer model parameters.

[0049] The intermediate layer aggregation unit 300 is responsible for performing intermediate layer model aggregation in federated learning, including multiple RSUs. After receiving the local model parameters from the local training unit 100, the intermediate layer aggregation unit 300 performs intermediate model aggregation on these parameters to obtain intermediate layer model parameters.

[0050] The global aggregation layer unit 400 is responsible for performing global model aggregation and training round reduction in federated learning, including the base station. After receiving the intermediate layer model parameters from the intermediate layer aggregation unit 300, the base station performs global model aggregation on these parameters to obtain global model parameters, and then sends them to the local training unit 100 through the intermediate layer aggregation unit 300. The training rounds are then adjusted based on the training results and broadcast to the intermediate layer aggregation unit and the local training unit to start the next round of model training.

[0051] Based on the aforementioned differential privacy-based federated learning system, a federated learning method for autonomous vehicle networking is provided, such as... Figure 2 As shown, the specific steps are as follows:

[0052] S100. Autonomous vehicles with local datasets perform the local training process of federated learning and obtain local model parameters.

[0053] Specifically, such as Figure 3 As shown, the specific steps are as follows:

[0054] Step 110: The local training unit 100 samples the sampling group to obtain the dataset sample set;

[0055] Specifically, with Sampling is performed based on the probability to obtain the dataset sample set. .

[0056] Step 120: Local training unit 100 processes each sample in the sample set obtained from sampling. Perform local gradient updates:

[0057]

[0058]

[0059] in, For the updated gradient, For dataset Size, This refers to the dataset samples. and local model parameters The loss function used for training can be convex or non-convex.

[0060] Step 130: Local training unit 100 prunes the updated gradients;

[0061] Specifically, because the gradients may become very large during network training, and the added dynamic programming noise may also distort the gradients, the local training unit 100 needs to clip the gradients after updating them.

[0062]

[0063] Where C is the gradient norm bound.

[0064] Step 140: Local training unit 100 updates the local model parameters based on the clipped gradients to obtain the updated local model parameters. ;

[0065] S200: The autonomous vehicle executes a privacy protection algorithm on the local model parameters and sends the parameters to the intermediate layer aggregation unit 300.

[0066] like Figure 4 As shown, the specific steps of the privacy protection unit 200 are as follows:

[0067] Step 210: Privacy protection unit 200 determines privacy parameters;

[0068] Specifically, given that federated learning needs to protect numerical data, a Gaussian mechanism is introduced to ensure... Differential privacy. Among them... Indicates the privacy budget, and indicates the current privacy level. It is the relaxation coefficient, which is defined. The failure probability of differential privacy.

[0069] Step 220: Privacy protection unit 200 calculates the sensitivity of the data that needs to be protected. ;

[0070] Step 230: Privacy protection unit 200 based on privacy parameters and sensitivity Calculations were performed to obtain a suitable Gaussian noise scale. ;

[0071] Step 240, the privacy protection unit 200 finally obtains the Gaussian noise scale. Add it to the original data to obtain the encrypted model parameters;

[0072] After the local iteration count meets the iteration conditions, the autonomous vehicle will proceed according to the privacy budget. and relaxation coefficient To local model parameters Add Gaussian noise, that is:

[0073]

[0074] Obtain the parameters of the local model containing noise. The data is then transmitted to the intermediate layer aggregation unit 300 via a wireless channel. If the conditions are not met, the local training unit 100 continues local training.

[0075] S300, each intermediate layer aggregation unit 300 performs intermediate aggregation on the received local model parameters to obtain intermediate model parameters;

[0076] S400 and each intermediate layer aggregation unit 300 perform privacy protection algorithms on the intermediate model parameters and send the parameters to the global aggregation layer unit 400;

[0077] Specifically, such as Figure 5 As shown, the training steps for each intermediate layer aggregation unit to perform the privacy-preserving algorithm on the intermediate model parameters are as follows:

[0078] S300: Each intermediate layer aggregation unit aggregates the received local model parameters to obtain intermediate model parameters. ;

[0079] Step 410: Determine if the intermediate layer iteration count has reached the intermediate layer iteration condition. If so, proceed to step 420A:

[0080] Step 420A. Each intermediate layer aggregation unit performs a privacy protection algorithm on the intermediate model parameters to obtain intermediate model parameters with added noise;

[0081] Specifically, the operation process of the privacy protection unit has been explained above. After obtaining the noise scale, the noise is added to the intermediate model parameters, that is...

[0082]

[0083] Obtain intermediate model parameters with added noise .

[0084] Step 430: Each intermediate layer aggregation unit compares the model benefit brought by training with the intermediate benefit threshold to obtain the number of RSUs whose training benefit is less than the threshold;

[0085] Specifically, This indicates that the loss on the test set is tested using the model parameters. This represents the benefit threshold of the intermediate layer model. The smaller the value, the less benefit the training brings to the model. If the following conditions are met:

[0086]

[0087] in This represents the number of training iterations for the local training layer. This represents the number of training iterations for the intermediate aggregation layer. This indicates that the benefit gained from training the intermediate layer model is less than the benefit threshold of the intermediate model.

[0088] Step 440: Each intermediate layer aggregation unit 300 sends the noisy intermediate model parameters and the number of RSUs with training benefits less than the threshold to the global aggregation layer 400 via a wireless channel.

[0089] If the intermediate layer iteration count does not reach the intermediate layer iteration count condition, then execute 420B:

[0090] Step 420B: Each intermediate layer aggregation unit broadcasts the aggregated intermediate model parameters to the autonomous vehicles within the region.

[0091] S500. Global aggregation layer unit 400 performs global aggregation on the incoming intermediate model parameters to obtain global parameters, and adjusts the number of training iterations based on the model benefits brought by training.

[0092] Specifically, such as Figure 6As shown, the training steps for the global aggregation layer unit are as follows:

[0093] Step 510: The global aggregation layer unit 400 aggregates the intermediate model parameters sent in to obtain global parameters.

[0094] Step 520: The global aggregation layer unit 400 reduces the number of intermediate aggregations based on the RSU ratio of the intermediate layer model training benefit being less than the intermediate model benefit threshold sent by the intermediate aggregation layer unit.

[0095] Specifically, when the benefits gained from training intermediate layers are less than the benefit threshold of the intermediate model. When the RSU ratio exceeds the benefit ratio γ, consider using Scale reduction of intermediate layer polymerization number The value of . Where t is the current round, This represents the number of times the data is aggregated globally. This is the attenuation coefficient, which should be adjusted according to the actual situation.

[0096] Step 530: The global aggregation layer unit determines the benefits of global training and the global model benefit threshold. The size relationship reduces the number of global training iterations;

[0097] Specifically, when the benefits of global training are less than the global model benefit threshold. At that time, consider using Scale reduction of global aggregation times The value of . Where This is the attenuation coefficient, which should be adjusted according to the actual situation.

[0098] Step 540: The global aggregation layer unit 400 sends the aggregated parameters to each autonomous vehicle through the intermediate aggregation layer unit 300.

[0099] Step 550: The global aggregation layer unit 400 broadcasts the new intermediate layer aggregation count to all intermediate layer aggregation units 300 and all autonomous vehicles. and global aggregation count ;

[0100] After step S500, there is also step S500A: determine whether the model performance meets the requirements or whether the number of iterations has reached the set value. If yes, the process ends; if no, return to step S100 to continue iterating the global model.

[0101] Specifically, if it is necessary to return to step S100 to continue the global model iteration, the global model aggregation unit 400 will use the intermediate model unit 300 to transfer the new round of global model parameters. The new intermediate layer aggregation count is then distributed to the autonomous vehicles. It is also broadcast to all intermediate layer aggregation units 300 and all autonomous vehicles. and global aggregation count .

[0102] The following is a specific example to illustrate the federated learning system and method for autonomous vehicle networking provided by this invention.

[0103] In the AVNs configuration, there are 100 AVs and 10 RSUs evenly distributed on the road, i.e., N = 100, M = 10. Each RSU has a coverage area of ​​10 vehicles, i.e., | =10. There is exactly one ECS server performing global aggregation.

[0104] First, the local training unit 100 samples the sampling group to obtain the dataset sample set, and sets the sampling group size. |=64, with Sampling is performed based on the probability to obtain the dataset sample set. Then, for each sample in the sampled sample set... The local gradient is updated to obtain the updated gradient. Finally, training unit 100 updates the local model parameters based on the clipped gradient to obtain the updated local model parameters. .

[0105] Secondly, the privacy protection unit 200 will update the local model parameters. Noise is added to obtain local model parameters containing noise. Then the autonomous vehicle transmits the encrypted local model parameters via wireless channel. Send to intermediate layer aggregation unit 300;

[0106] Then the intermediate aggregation unit 300 will process the sent local model parameters. Intermediate model parameters are obtained by aggregating intermediate layer models. Then, the privacy protection unit 200 is used to process the intermediate layer model parameters. Noise-adding processing is performed to obtain the intermediate layer model parameters containing noise. After the intermediate layer iteration conditions are met, the intermediate layer aggregation unit 300 transmits the noisy intermediate layer model parameters via a wireless channel. Send to global aggregation layer unit 400.

[0107] Finally, the global aggregation layer unit 400 sent the intermediate layer model parameters. Perform global aggregation to obtain global model parameters The number of training iterations for the intermediate layers is adjusted based on the training benefits and benefit thresholds at this point. Number of training iterations for the global aggregation layer .

[0108] Once the final model performance meets the requirements, the local model parameters, intermediate model parameters, and global model parameters will be output, along with the model's accuracy, loss function value, and total number of communication rounds. The changes.

[0109] The final model performance report shows that the accuracy of the federated learning system for autonomous vehicle networking proposed in this invention can reach over 95%, and its performance in the early stages of iteration is superior to the training method without added noise. Since the noise added by AV and RSU both follow a Gaussian distribution, the noise partially cancels each other out, thus weakening its impact on model performance. Therefore, under the current privacy budget parameter settings, both privacy and performance are protected. Furthermore, the results show that the total number of communication rounds in this invention decreases rapidly in the later stages of training because the impact of noise becomes significant in the later stages. Continuing training under these circumstances would actually reduce model performance, thus decreasing the total number of communication rounds. T It will decline relatively quickly, ending training early.

[0110] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A federated learning system for autonomous vehicle networking, characterized in that, It includes a local training unit, a privacy protection unit, an intermediate layer aggregation unit, and a global aggregation layer unit; wherein, the local training unit executes the privacy protection unit, the local training unit is connected to the intermediate layer aggregation unit via a wireless channel, the intermediate layer aggregation unit executes the privacy protection unit, and the intermediate layer aggregation unit is connected to the global aggregation layer unit via a wireless channel; The local training unit is responsible for performing local training in federated learning, including autonomous vehicles participating in federated learning; the privacy protection unit is responsible for protecting the privacy of the local model parameters of the local training unit and the intermediate model parameters of the intermediate layer aggregation unit. The intermediate layer aggregation unit is responsible for performing intermediate layer model aggregation in federated learning, including the drive test unit; the global aggregation layer is responsible for performing global model aggregation in federated learning and reducing the number of intermediate layer aggregations and global aggregations, including the base station. The privacy protection unit is implemented using a differential privacy algorithm, and the local training unit and the intermediate layer aggregation unit also execute the differential privacy algorithm to ensure that the federated learning system has higher security. The global aggregation layer unit reduces the number of intermediate layer aggregations and global aggregations based on training efficiency, and broadcasts this information to the intermediate layer aggregation unit and the local training unit to ensure training efficiency.

2. A federated learning method for autonomous vehicle networking, characterized in that, Includes the following steps: S100. Autonomous vehicles containing local datasets perform local training in federated learning to obtain local model parameters; S200: The autonomous vehicle executes a privacy protection algorithm on the local model parameters and sends the noisy local model parameters to the intermediate layer aggregation unit via a wireless channel; S300: Each intermediate layer aggregation unit performs intermediate layer training in federated learning on the received local model parameters to obtain an intermediate model; S400, Each intermediate layer aggregation unit performs a privacy protection algorithm on the intermediate model parameters and sends the parameters to the global aggregation layer unit via a wireless channel; specifically including the following steps: Step 410: Obtain the intermediate layer model parameters; determine if the number of iterations has reached the intermediate iteration count condition. If yes, proceed to step 420A; otherwise, proceed to step 420B. Step 420A: Each intermediate aggregation layer unit broadcasts the aggregated parameters to the autonomous vehicles in the area; Step 420B: Each intermediate layer performs a privacy protection algorithm on the intermediate layer model parameters to obtain intermediate layer model parameters with added noise; Step 430: Compare the benefits of training each intermediate layer with the intermediate model benefit threshold to obtain the number of RSUs whose training benefits are less than the threshold; Step 440: Send the parameters of the noisy intermediate layer model and the number of RSUs with training benefits less than the threshold to the global aggregation layer unit via a wireless channel; The S500 global aggregation layer performs global aggregation in federated learning on the received intermediate model parameters to obtain global parameters. Based on the training efficiency, it adjusts the training epochs and outputs all model parameters and the number of training epochs at this time.

3. The federated learning method for autonomous vehicle networking as described in claim 2, characterized in that, The process after step S500 includes S500A: determining whether the model performance meets the requirements or whether the number of iterations exceeds the set value. If yes, the process ends; otherwise, it returns to step S100 to continue iterating the global model.

4. The federated learning method for autonomous vehicle networking as described in claim 2, characterized in that, S100. Autonomous vehicles containing local datasets perform local training in federated learning to obtain local model parameters, specifically including the following steps: Step 110: The local training unit of the autonomous vehicle samples the sample group to obtain the dataset sample set; Step 120: The local training unit updates each sample in the obtained sample set locally; Step 130: The local training unit prunes the gradient of each updated sample; Step 140: The local training unit updates the local model based on the local gradient obtained after pruning; then, it is determined whether the number of iterations has reached the local iteration count condition. If yes, then proceed to step 150; otherwise, return to step 110 to continue local model training. Step 150: The local training unit performs a privacy protection unit on the local model parameters to obtain local model parameters with added noise.

5. The federated learning method for autonomous vehicle networking as described in claim 4, characterized in that, The S500 global aggregation layer performs global aggregation in federated learning on the received intermediate model parameters to obtain global parameters, and adjusts the training epochs based on the training efficiency. Specifically, it includes the following steps: Step 510: The global aggregation layer unit aggregates the intermediate layer model parameters sent in to obtain the global model parameters; Step 520: The global aggregation layer unit reduces the number of aggregations in the intermediate aggregation layer based on the RSU ratio of the intermediate layer training benefit that is less than the threshold sent by the intermediate layer aggregation unit. Step 530: The global aggregation layer unit determines the relationship between the benefits of global training and the global model benefit threshold, thereby reducing the number of global aggregations. Step 540: The global aggregation layer unit broadcasts the aggregated parameters to each autonomous vehicle; Step 550: The global aggregation layer unit broadcasts the new aggregation count of the intermediate layer aggregation unit and the global aggregation layer to all intermediate layer aggregation units and all autonomous vehicles.

Images